This repository, NVlabs/tiny-cuda-nn, is a high-performance C++/CUDA library and PyTorch extension for training and querying neural networks with a focus on extreme GPU efficiency. It provides a modular, JSON-configurable framework of specialized components: input encodings (multiresolution hash grids, spherical harmonics, frequency, etc.), network architectures (fully fused MLP using Tensor Cores, CUTLASS-based MLP), loss functions, and optimizers (Adam, Shampoo, Novograd, etc.). The library supports runtime JIT compilation to fuse forward, backward, and loss kernels into single CUDA kernels, yielding up to 5x speedups. It is the engine behind NVIDIA's Instant Neural Graphics Primitives (Instant NGP) and is widely used in NeRF and neural rendering research. The codebase is production‑quality, with extensive use of templates, GPU memory arenas, CUDA graphs, multi‑stream parallelism, and comprehensive CI across many CUDA versions and GPU architectures. Although primarily a library, it includes sample applications and benchmarks that demonstrate its use in image overfitting and throughput measurement. PyTorch bindings via pybind11 make it accessible for Python‑based research workflows.
sunnypilot is a community fork of comma.ai's openpilot, an open-source driver assistance system (ADAS) designed for real-time operation on dedicated hardware (comma three/4, panda). The codebase implements the full autonomy stack: camera and radar perception, path planning, longitudinal/lateral control, driver monitoring, and localization via Kalman filters. It runs as a set of daemon processes orchestrated by a process manager, communicating through a shared-memory pub/sub messaging layer (msgq) with Cap'n Proto serialization. The system includes a full-featured Raylib-based UI with on-road HUD, settings panels, and an installer, plus cloud connectivity through Athena (remote commands, registration) and the fork-specific Sunnylink (settings backup/restore, telemetry). Beyond the core openpilot functionality, sunnypilot adds numerous extended features: Modular Assistive Driving System (MADS) for configurable steering modes, neural network lateral control (NNLC), map-based speed limit assist using OpenStreetMap data (mapd), intelligent cruise button management, dynamic experimental control, and a model manager that fetches and caches driving model bundles from a remote manifest. The fork maintains its own cereal schema extensions and custom UI layouts for both touch (mici) and big-screen (tizi) hardware variants. The repository includes extensive testing infrastructure: unit tests for most modules, process replay tests that compare outputs against reference logs, longitudinal/lateral maneuver simulations, and CAN bus analysis tools (cabana, replay). It also provides simulation bridges (MetaDrive, joystick) for closed-loop testing. The build system uses SCons for C++ components and uv for Python dependencies, with CI pipelines for model building, release drafting, and route uploads.
Mastodon is a full-stack, federated social network server implementing the ActivityPub protocol. It provides a production-grade Rails API backend with a React/Redux frontend, real-time streaming via Node.js, and extensive background job processing with Sidekiq. The system supports user timelines, media attachments, content moderation, admin dashboards, OAuth authentication, and federation with other ActivityPub-compatible services. The codebase is large and actively maintained, with a clear separation of concerns across controllers, services, serializers, workers, and frontend components. It includes advanced features such as trending content, annual user reports, FASP data sharing, and a comprehensive admin interface. While the architecture is well-structured, there are signs of technical drift including mixed frontend patterns (legacy JSX vs modern TypeScript/Redux Toolkit), placeholder Ruby/Node version numbers, and some deprecated API endpoints.
VieNeu-TTS is an advanced on-device Vietnamese text-to-speech library with instant voice cloning, supporting bilingual (Vietnamese/English) synthesis. The system provides a Python SDK (`vieneu`) with multiple backends: v3 Turbo (default, 48kHz, ONNX or PyTorch), v3 Nano (lightweight 24kHz, ONNX only), and legacy v1/v2 (deprecated). It includes a Gradio web UI (`vieneu-web`), a FastAPI streaming server, and a browser-based HTML/JS client. The core architecture is a factory pattern (`Vieneu(mode=...)`) that returns the appropriate engine, all inheriting from a shared base class. Inference can run entirely torch-free on CPU via ONNX Runtime, with optional GPU acceleration (CUDA) for batched generation and lower latency. Feature highlights include instant voice cloning from a 3–8 second reference clip using a speaker encoder (ONNX) and denoiser, inline emotion cues (`[cười]`, `[thở dài]`), frame-level streaming, SRT subtitle-to-speech, and LoRA fine-tuning for custom voices (requires ~6GB GPU). The system is deployed via Docker with CPU and GPU profiles, and has CI/CD pipelines for testing and PyPI publishing. The codebase is well-structured with clear separation of concerns, though it carries some legacy weight from older model versions and a few deprecated parameters. Other notable components include a sliding-window repetition penalty, a babble guard heuristic for short utterances, and a voice persistence system for user-defined presets.
DEIMv2 is a production-grade object detection framework built upon the DETR architecture, specifically implementing the DEIM (DETR with Improved Matching) algorithm extended with DINOv3 features. The codebase provides a modular design with multiple backbone options (HGNetv2, DINOv3, ResNet variants), hybrid encoders, and transformer decoders (DEIM, D-FINE, RT-DETRv2). It includes comprehensive training scripts, YAML-based configuration inheritance, data augmentation pipelines, and deployment tools for ONNX, TensorRT, and OpenVINO. The repository directly trains and evaluates on standard detection benchmarks like COCO and Objects365, and is structured for reproducibility and extensibility. While heavily inspired by prior works (RT-DETR, D-FINE), the implementation demonstrates solid engineering with clean abstractions, distributed training support, and a well-organized module hierarchy.
Trice is an embedded logging and tracing toolchain consisting of a lightweight C instrumentation library for microcontrollers and a Go-based host CLI tool. The C library provides variadic macros that emit compact binary records (ID + values) instead of format strings, drastically reducing flash and bandwidth usage. The Go tool manages Trice IDs (insert, clean, bind), decodes incoming binary streams from targets over UART, RTT, TCP/UDP, or file, and displays formatted logs with timestamps, colors, and filtering. It supports multiple buffer strategies (stack, static, double, ring), framing protocols (COBS, TCOBS), optional XTEA encryption, and local logging on the target itself. The architecture is a classic host-target split: the target runs the C library, the host runs the Go CLI. The codebase is well-tested with a large matrix of CGO test configurations and CI pipelines for multiple STM32 example projects. It is actively maintained with frequent releases and a responsive issue tracker.
Hasura GraphQL Engine is an open-source, metadata-driven GraphQL server that provides instant, real-time GraphQL APIs over multiple databases (PostgreSQL, MSSQL, BigQuery, CockroachDB, Citus, and custom sources via Data Connectors). The repository is a large monorepo containing a Rust-based v3 engine, a Go CLI for project management (migrations, metadata, seeds), a React/TypeScript admin console, and community boilerplates for authentication, event triggers, and remote schemas. The system exposes a single GraphQL endpoint that supports queries, mutations, subscriptions, streaming subscriptions, event triggers, cron triggers, remote joins, actions, and Apollo Federation. It relies on external databases, webhook endpoints for auth and event handling, and optionally OpenTelemetry collectors for observability. The architecture is modular: the v3 engine handles metadata resolution, GraphQL schema generation, and execution planning; the CLI manages project lifecycle; and the console provides a GUI for configuration. The codebase is actively maintained with frequent releases, but shows signs of technical drift due to legacy frontend code (Redux vs React Query), dual version support (v2/v3), and multiple programming languages (Rust, Go, TypeScript, Haskell, Python).
This repository is a collection of deployable templates and educational cookbooks for building real-time LLM applications using the Pathway framework. It provides ready-to-deploy RAG (Retrieval-Augmented Generation) solutions with live data synchronization from various sources (files, Google Drive, SharePoint, etc.), built-in vector and hybrid search, and integration with LangGraph for agentic workflows. The templates are production-oriented with Docker support, while the cookbooks demonstrate advanced patterns like Self-RAG with self-reflection and query rewriting. The codebase is well-structured but exhibits some duplication across templates, and the core logic relies heavily on the external Pathway library.
Meteor is an ultra-simple full-stack JavaScript framework for building modern web and mobile applications. This monorepo contains the Meteor CLI tool (`meteor`), the build system (isobuild, Rspack integration), core runtime packages (reactive data via DDP, MongoDB integration, accounts, Blaze templating, modules, bundling), and a rich ecosystem of OAuth providers, testing infrastructure, and developer tools (ESLint plugin, Babel preset, installer). The codebase is organized as a monorepo with `tools/` for the CLI and build system, `packages/` for framework packages, and `npm-packages/` for standalone npm packages. It demonstrates a real-time data synchronization pattern using DDP (Distributed Data Protocol), a custom constraint solver for package dependency resolution, and a compiler plugin system for transpilation. Despite being a well-established framework with extensive testing, the codebase shows signs of technical drift: deprecated packages remain in-tree, there is a dual sync/async API surface due to ongoing migration from Fibers to async/await, and the build system is transitioning from a legacy compiler to Rspack.
Zulip is a production-grade, open-source team chat application built as a Django monolith with a real-time Tornado event server. The backend uses Django ORM with PostgreSQL, RabbitMQ for async queue processing (email, push notifications, URL previews), Memcached/Redis for caching and rate limiting, and a custom typed-endpoint decorator pattern for API validation. The frontend is a mix of legacy jQuery/Handlebars and modern TypeScript with Zod schemas, Webpack bundling, and extensive UI components for chat composition, streams, topics, user groups, and settings. The system exposes a REST API (used by mobile and desktop clients), webhook ingestion for 50+ third-party services, and a full billing/support subsystem via Stripe. The architecture is a single monolithic Django application with separate Tornado process for real-time push, alongside queue workers and cron-like management commands. Key entry points include the Django WSGI server, Tornado event server, management commands (e.g., manage.py), and webhook view functions. External dependencies include PostgreSQL, RabbitMQ, Redis, Memcached, Stripe, pgroonga (full-text search), pyvips (image processing), and numerous npm packages.
Feldera is a real-time streaming SQL analytics platform built on the DBSP (Differential Dataflow) incremental computation engine. It compiles SQL into efficient incremental circuits that process streaming data with sub-millisecond latency, supporting complex queries including joins, aggregations, window functions, and recursive queries. The system is deployed as a pipeline manager (Rust) with a REST API, a web console (Svelte), a Python SDK, a dbt adapter, and a CLI tool. It integrates with numerous data sources and sinks (Kafka, Postgres, Delta Lake, Iceberg, HTTP, NATS, Redis, etc.) and provides fault tolerance via checkpointing and storage backends. The architecture is polyglot: a Java-based SQL compiler (using Apache Calcite) translates SQL to Rust code that runs on the DBSP runtime; the pipeline manager orchestrates compilation, deployment, and lifecycle; the web console offers a full UI for pipeline management, monitoring, and profiling. The codebase also includes a Spark SQL translator (felderize) and extensive benchmark suites (Nexmark, LDBC). While primarily an application, it has strong research roots in incremental view maintenance and formal monotonicity analysis, reflected in the compiler's optimization passes and the DBSP runtime's operator design. Key reusable components include the adapter library (connector traits), the Feldera types crate (shared configs and errors), the common UI component library, and the Python client SDK. The project is actively maintained with frequent commits, comprehensive CI/CD, and a growing test suite covering SQL semantics across many data types and edge cases.
Iodine is a high-performance concurrent web server for Ruby applications, built as a wrapper around the C-based facil.io framework. It provides native HTTP/WebSocket/Server-Sent Events support, Rack compliance for serving Ruby web applications, and a built-in Pub/Sub system with optional Redis scaling for horizontal distribution. The architecture follows a master-worker model using multiple processes and threads to handle thousands of concurrent connections, with critical networking components implemented in C for performance and Ruby bindings for developer productivity. Key features include static file serving with gzip and X-Sendfile support, a secure Mustache template engine with automatic HTML escaping, a fast JSON parser claiming 20-90% speed improvements over Ruby's standard library, and TLS support (though marked as having known issues for production use). The server integrates with Rack applications through a custom handler and provides upgrade hooks for WebSocket/SSE protocols via the Rack environment. Pub/Sub functionality uses an engine pattern allowing swap between internal cluster-based and external Redis backends. Entry points include the `exe/iodine` CLI executable and programmatic controls via `Iodine.listen`, `Iodine.start`, and `Iodine.stop`. Data flows involve HTTP requests parsed by the C-based HTTP processor, routed to Rack applications or static file handlers, with WebSocket connections upgraded via Rack's `env['rack.upgrade?']` mechanism. The Pub/Sub system operates independently for real-time messaging across processes or machines. External dependencies include the vendored facil.io C library, OpenSSL development headers for TLS support, and optionally the Redis gem for clustered Pub/Sub.
Aerugo is a safety-critical real-time operating system (RTOS) for embedded systems, specifically targeting the ARM Cortex-M7 (ATSAMV71) microcontroller. Developed under an ESA contract for space applications, it provides a non-preemptive, cooperative scheduler based on fine-grained tasklets, along with event-driven and cyclic execution, message queues, boolean conditions, and execution monitoring. The system is designed for static allocation, no dynamic memory, and no_std environments. The repository includes a full hardware abstraction layer (HAL) for the SAMV71, covering SPI, UART, PIO, timers, watchdog, XDMAC, PMC, FPU, and more, built on top of an auto-generated peripheral access crate (PAC) from the vendor SVD. A custom Python test framework (calldwell) enables hardware-in-the-loop testing via RTT, GDB, and SSH, with formal requirements traceability embedded in test code. The architecture is modular: the core RTOS kernel resides in `src/`, platform-specific code in `arch/` (Cortex-M and x86), and reusable utilities in `utils/` (bitfield-enum macro, bounded integers, LSM6DSO IMU driver). The HAL uses typestate patterns to enforce correct driver configuration at compile time. The codebase demonstrates strong Rust embedded engineering practices, including manual vtables for type erasure, critical-section-based mutexes, and static resource management via OnceCell. However, it relies on nightly Rust features (const_mut_refs, proc_macro_tracked_env) and the PAC is auto-generated, introducing maintenance risks. The project is actively developed with a comprehensive test suite and real-world demos (e.g., SPI accelerometer with CCSDS telemetry).
Simple Open EtherCAT Master (SOEM) is a lightweight, portable C library that implements the EtherCAT master protocol stack. It provides full support for CoE (CANopen over EtherCAT), FoE (File over EtherCAT), SoE (Servo over EtherCAT), EoE (Ethernet over EtherCAT), Distributed Clocks (DC), SII (Slave Information Interface), and network redundancy (dual NIC). The architecture is modular: an OS abstraction layer (osal) encapsulates platform-specific threading, timing, and memory operations, while a hardware abstraction layer (oshw) handles raw network access via sockets (Linux, Windows, RTEMS, VxWorks, etc.). The core library functions (ecx_*) operate on a context structure, enabling multiple master instances. Key implementation patterns include an indexed buffer management system for out-of-order frame handling, a thread-safe mailbox pool and queue, and a circular error ring buffer. The library is not a standalone application; it is designed to be embedded into industrial real-time control systems. Sample applications demonstrate cyclic process data exchange, SDO access, EEPROM manipulation, and firmware updates. The codebase is mature (2005–2025) and under active development, with a dual GPLv3/commercial license. Some technical drift exists due to bundled third-party WinPcap headers and several platform-specific contributor directories that may lag behind core updates. Overall, the library demonstrates strong engineering with clear separation of concerns and portability, making it a solid foundation for EtherCAT master development.
This repository is the official JavaScript/TypeScript client SDK for Centrifugo, a scalable real-time messaging server. It provides a robust, event-driven client (Centrifuge class) that manages persistent connections via multiple transports (WebSocket, SSE, HTTP stream, SockJS, WebTransport), handles token-based authentication and subscription authorization, and supports three subscription models: stream (standard channel), map (keyed data), and shared poll (polling with version tracking). The SDK also implements advanced features like automatic reconnection with exponential backoff, recovery from connection drops using offset/epoch, command batching for HTTP transports, and delta compression for efficient updates. The architecture is library-oriented with a clear separation of concerns: transport implementations are abstracted behind a common interface, the main client class manages connection state and dispatch ordering via a serial promise chain, and each subscription type extends a base class with its own state machine (stream, map three-phase, shared poll). The codebase is extensively tested with both unit tests (using a fake server) and integration tests against a real Centrifugo instance in Docker. The SDK is published on npm, supports CJS and ESM builds, and includes protobuf codec support. Notable patterns include a transport abstraction with dependency injection that allows easy addition of new transports, a serial dispatch chain ensuring ordered response processing, and a per-key publish debounce mechanism. The library is actively maintained with ongoing work on refactoring subscription types as documented in NEXT_MAJOR_SUBSCRIPTION_TYPES.md. The repository also demonstrates high craftsmanship in its comprehensive test coverage, graceful error handling, and careful handling of edge cases like stale tokens, network interruptions, and concurrent operations.
Yave is a C++20 Vulkan-based real-time graphics engine paired with a custom ImGui-driven editor and a self-contained core utility library named `y`. The repository is structured as three layers: `y` provides reusable containers, math, concurrency, IO, serialization, and a test framework; `yave` implements the engine itself (Vulkan device/window management, frame graph, render passes, ECS, asset pipeline, physics, animation, and scripting); and `editor` is a full 3D scene editor application with workspaces, widgets, gizmos, importers, thumbnails, and undo/redo. The renderer supports deferred PBR, cascaded shadow maps, TAA, bloom, SSAO/RTAO, RTGI, atmosphere scattering, and tone mapping, and it runs through a declarative frame graph with automatic resource pooling and barrier insertion. The editor is not a separate toy: it provides asset browsing, glTF/image import, material editing, scene outliner, inspector, file system views, performance metrics, GPU profiling, and RenderDoc integration. The engine uses a modern ECS with component groups, async asset pointers, and a serialization framework that supports schema evolution. The `y` library is particularly notable for its clean, high-performance primitives: SSO strings, flat hash maps with SIMD probing, sparse component sets, dependency-aware job system, and a forward/backward-compatible binary serialization layer. Externally, it depends on Vulkan, VMA, Volk, Dear ImGui, JoltPhysics, LuaJIT, tinygltf, stb, Tracy, and spirv_reflect, mostly vendored. While the project is described by its author as a learning/pet engine, the code is generally well-structured and professionally organized. However, it is a single-author codebase with many `Y_TODO` markers, some stubbed platform paths, global device singletons, and an explicit compiler-dependent serialization format. Tests exist mainly for the `y` core containers and math, not for the engine or editor. The repository is active and uses current C++/Vulkan practices, but should be treated as an evolving research-quality engine, not a stable library.
FUXA is a full-stack web-based SCADA/HMI builder and runtime for industrial automation and IoT. It provides a visual editor (Angular + SVG) to create real-time dashboards, connect to field devices via 13+ protocols (Modbus, OPC UA, BACnet, Siemens S7, MQTT, etc.), and manage alarms, data acquisition (DAQ), scripting, scheduling, and user permissions. The backend (Node.js/Express) communicates with the frontend via Socket.IO for real-time device value updates. The system supports multiple time-series databases (SQLite, InfluxDB, QuestDB, TDengine) and can be deployed as a web server, Docker container, or Electron desktop app. It also integrates with Node-RED for low-code automation flows. The architecture is modular with a device driver factory pattern, JWT-based authentication, and a plugin system for extensibility. While the codebase is large and contains some monolithic components, it demonstrates solid engineering practices and has a rich set of reusable UI components and utility patterns.
Helios is a real-time long video generation system that implements text-to-video (T2V), image-to-video (I2V), and video-to-video (V2V) generation using a 3D transformer architecture. The system features a modular design with specialized optimizations for performance including Ahead-of-Time Compilation (AOTI), group offloading for VRAM reduction, context parallelism (Ulysses/Ring attention), and Flash Attention kernel optimizations. It employs a multi-stage training pipeline (Base → Mid → Distilled) to balance quality and sampling speed. The repository includes a Gradio web interface (app.py) for interactive generation, command-line inference scripts, training utilities, and a comprehensive evaluation suite measuring video quality across aesthetic, motion, semantic, and naturalness dimensions. While containing research elements for ablation studies and benchmarking, its primary function is as a deployable video generation application.
FluxVLA is a comprehensive, config-driven framework for training, evaluating, and deploying Vision-Language-Action (VLA) models for robotic manipulation. It supports multiple VLA architectures (OpenVLA, Pi0/Pi0.5, GR00T, Cosmos3, DreamZero, etc.) and robot platforms (LIBERO simulation, RoboCasa, Franka, UR3, ALOHA, Oli humanoid, Tron2). The codebase provides a modular pipeline with registries and builders for datasets, transforms, models (backbones, heads, projectors), optimizers, collators, and inference runners. Production capabilities include distributed training (FSDP/DDP), real-time ROS inference, custom CUDA/Triton kernels for performance, and evaluation with metrics reporting to Feishu. While primarily an engineering framework for applied robot learning, it also contains research contributions such as flow-matching action heads, reward models (ARM, SARM), and real-time chunking (RTC) guidance. The architecture is clean and extensible, though config proliferation and vendored third-party code introduce maintenance overhead.
AppFlowy-Web is a feature-rich collaborative workspace application built with React, TypeScript, and Yjs (CRDT). It provides real-time collaborative document editing via a Slate-based editor, multiple database views (grid, board, calendar, gallery, list, timeline, feed, chart), AI-powered writing assistance and chat, publishing/sharing capabilities, and workspace management. The application is designed as an open-source alternative to Notion, supporting both cloud-hosted and self-hosted deployments. The architecture is a single-page application with Yjs as the core data model for offline-first collaboration, backed by IndexedDB caching and REST/WebSocket APIs. The codebase is extensive (4000+ files) with comprehensive unit and E2E tests, but exhibits moderate technical drift due to mixed UI paradigms (MUI + Radix), complex Yjs observer lifecycle management, and some deprecated environment variable mappings.